News · 2026-07-22

AI Revenue Now Covers the Data-Center Depreciation Bill, But Not the Full Cost

For the first time, modeled AI revenue has crept past the cost of depreciating the hardware it runs on, according to an industry report, but the margin is thin and the milestone is far narrower than "AI is now profitable." The report finds that late 2025 was the first quarter in which estimated AI revenue exceeded estimated AI-infrastructure depreciation, with a modest cushion that shrinks or vanishes depending on how long you assume the chips stay useful. It is a real, specific data point in the capex debate, and it is being over-read in both directions.

Key facts

• Exponential View's June 2026 State of the AI Economy report models more than 1,000 firms with confidence-scored inputs.
• It finds Q4 2025 was the first quarter modeled AI revenue exceeded modeled AI-infrastructure depreciation.
• In Q1 2026, depreciation absorbed roughly 81 percent of hyperscaler and neocloud generative-AI revenue, leaving about 19 percent headroom before other costs.
• Primary source: the report PDF.

The background is the central anxiety of the current AI boom: hyperscalers and specialized cloud providers are spending staggering sums on GPUs and data centers, and the question is whether the revenue AI generates can ever justify that outlay. Depreciation is the accounting way of spreading a hardware purchase across its useful life, so a fleet of chips expected to last six years shows up as one-sixth of its cost as an expense each year. Asking whether revenue covers depreciation is asking a deliberately modest question, is the business at least earning back the annual wear-and-tear on its equipment, before you even count electricity, staff, and the original financing.

The report's narrow finding is that the answer flipped to yes in the fourth quarter of 2025. In the first quarter of 2026, depreciation absorbed about 81 percent of the modeled generative-AI revenue at hyperscalers and neoclouds, leaving roughly 19 percent of headroom, and the report is explicit that this calculation excludes operating expenses. Its separate illustrative data-center model does include energy, staff, maintenance, overhead, cost of capital, and in one scenario model licensing, which is a reminder that clearing the depreciation bar is not the same as clearing the total-cost bar.

The fragility of the result is the most important caveat. The whole thing hinges on assuming a six-year useful life for the IT equipment. The report shows that shorter chip lives eliminate the modeled revenue coverage entirely, while longer lives increase the headroom. Given how fast AI accelerators are being superseded, whether a GPU bought today is genuinely productive in 2032 is a live question, and the answer swings the conclusion from "barely covered" to "not covered."

There is also a genuinely useful second finding about demand. The report estimates token-demand elasticity of roughly 1.2 to 1.8, meaning a 10 percent cut in price corresponds with 12 to 18 percent more tokens consumed. If that holds, total spending on AI can actually rise as unit prices fall, because cheaper tokens get used more than proportionally. The authors caution this is a time-series relationship that may overstate pure price sensitivity, but it is the kind of dynamic that explains why falling per-token prices have not translated into falling AI bills.

Why it matters: this is one of the few attempts to put a number on the "does the AI economy add up" question using a consistent model rather than vibes. It offers the capex bulls a qualified win, revenue now covers the annual depreciation meter, and hands the bears an equally valid rejoinder, only barely, only excluding operating costs, and only if the hardware lasts six years. It pairs naturally with the parallel story about metered access, where the viral claim that the US Army "exhausted its unlimited AI tokens" is unverified, even though Army documents do confirm its Ask Sage access is token-metered at 200,000 free tokens per user per month.

The honest caveat: these are modeled estimates, not audited accounts. The report draws on confidence-scored inputs that include executive comments, proxies, unverified estimates, and leaks, and it works hard to avoid double-counting revenue across the application, model, and hosting layers, but it remains an estimate. The reception among analysts has been assumption-skepticism, questioning the six-year service life and asking for actual-versus-predicted depreciation, rather than independent validation. Read it as the best available modeled snapshot of a genuinely uncertain picture, not as a verdict.

Primary source, verified: read the paper →

Ground Truth. (2026, July 22). AI Revenue Now Covers the Data-Center Depreciation Bill, But Not the Full Cost. Ground Truth. https://groundtruth.day/news/ai-revenue-finally-covers-the-datacenter-depreciation-bill.html

@misc{groundtruth:ai-revenue-finally-covers-the-datacenter-depreciation-bill,
  title  = {AI Revenue Now Covers the Data-Center Depreciation Bill, But Not the Full Cost},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/ai-revenue-finally-covers-the-datacenter-depreciation-bill.html}
}

Topics: ai-economics · infrastructure · capex · industry